AI Doesn't Guess, It Fills in the Gaps (And That's Why It Hallucinates)
AI Doesn't Guess, It Fills in the Gaps (And That's Why It Hallucinates)
When AI generates "generic responses", we need to understand what lies behind the models. Only then can we optimize their results and gain real value from this technology.
You have probably experienced this already.
You ask AI for something, it replies instantly, the text looks great, well-organized, full of bullet points.
Except it's completely useless.
The answer is "technically correct", but totally generic.
You read it, think "I could write this much better myself", close the window, and do it manually.
The natural conclusion is that the tool isn't mature yet, or that your use case is too specific for it.
In reality, the problem is entirely different.
In one of the corporate AI training cohorts AutoU has been delivering inside enterprise companies, Alexandre Canellas, our technical lead, demonstrated live why this happens.
And the most fascinating part was watching the exact same model, in the exact same session, deliver a useless response first, and right after that, a response ready to send.
The difference wasn't in the tool, nor in the underlying model.
By the end of this article, you will understand:
* why AI responds generically, and why hallucination is the exact same mechanism taken one step further;
* the four-block prompt framework that transformed the live response, using the actual prompt from the session;
* the golden rule to know when context is worth providing and when it's just a waste of time.
1. The Test
Alexandre opened Gemini and deliberately typed the most generic request possible:
“Write me a standard quote email for a client.”
The AI returned a full email in seconds. Subject: “Proposal Quote”, polite opening, detail sections, summary, next steps. All neatly bulleted.
No one in the room would ever send that email to an actual client.
Then he asked the question that cuts through the confusion: What did the AI have to work with?
He hadn't stated what was being quoted, to whom, in what tone, or in what format he wanted the reply.
2. What AI Does with What You Left Unsaid
The mechanism is simple, and it's worth remembering because it explains almost everything that follows.
AI never generates information out of thin air. It always generates based on the reference material it has.
When you provide no context, it doesn't pause to ask questions. It fills the gap with whatever reference it holds—which, in a generalist foundation model, is the statistical average of everything it has ever seen on the internet on that topic.
In the case of that email, its reference consisted of thousands of sales quote templates, "how to write a business email" articles, and canned boilerplates. It gathered all of that, extracted the pattern, and returned the most statistically probable outcome.
And the most probable outcome, by definition, is the generic average.
→ A generic response isn't a flaw of the model; it is the internet's average rushing in to fill the vacuum you left behind.
A telltale sign Alexandre pointed out, which I had never noticed before: excessive bullet points are a symptom of this. When an LLM lacks reference, it defaults to bulleted lists. When it has rich context, it defaults to natural flowing prose tailored to your voice.
3. Hallucination Is the Same Mechanism, Just Taken Further
A participant in the cohort asked, with good reason, what hallucination really was. Alexandre's explanation was far simpler than expected.
Hallucination happens when the gap you leave is too wide, and the AI, in attempting to fill it, starts fabricating another narrative entirely.
The example he used was an itinerary. You ask for a travel itinerary thinking of a business conference, but the AI doesn't know that and drafts a tourist vacation itinerary. It touches upon your request, but misses the mark completely.
And there is a compounding issue that few people notice, because it creeps in gradually.
When you attempt to correct it, your correction enters the conversation alongside the error. The context window now carries your prompt, the flawed answer, and your correction. There comes a point where the AI no longer has clean references to proceed, and the thread becomes unsalvageable.
At that stage, the fastest route is opening a fresh chat and specifying constraints properly from the start. Persisting in a corrupted thread almost always makes it worse.
4. The Structure That Fixed It
Right after that, Alexandre opened a new chat in the same model and reformulated the prompt into four blocks. Here is the exact prompt typed during the session:
Persona: You are the sales manager at an electrical engineering company, maintaining a professional yet approachable tone with clients.
Context: The client requested a quote for a photovoltaic solar energy installation in a 500 m² commercial warehouse. We sent the technical proposal last week and haven't heard back yet.
Instruction: Write a follow-up email asking if they have any questions regarding the proposal and reminding them of the quote's validity window.
Format: Short email, under 100 words, including a suggested subject line.
The response returned in natural flowing prose, with a relevant follow-up subject, mentioning the specific project, inquiring about open questions, and noting that proposal terms remain valid for 30 days.
It was ready to send with just one or two minor adjustments.
Same model, same session. The only variable that changed was the input material provided to the AI.
Of the four blocks, context is by far the most decisive. It is what drags the AI away from statistical internet averages and anchors it firmly in your reality.
Format is the block people most frequently skip. If you don't dictate how you want the answer structured, the model decides for you. At AutoU, we've had enterprise projects where format mattered just as much as context, because the response was feeding directly into an internal system and required strict structured JSON rather than free text.
5. The Trick to Never Forget the Framework
Alexandre shared a rule of thumb that removes the need to memorize templates.
Before writing the prompt, put yourself in the AI's shoes.
If a colleague walked over to your desk and said "write me a quote email", what would you ask back?
Quote for what? To whom? New or returning client? Formal or casual? Brief or detailed?
Those questions are your prompt. The difference is that you answer them upfront, rather than letting the AI guess.
6. “I'm Too Lazy to Provide So Much Instruction”
The most candid remark of the class came from a sales executive:
“Honestly speaking, I feel a bit lazy having to write so many instructions just to draft an email. Usually I just write my own draft and ask the AI to polish it.”
And she is right—at least in part.
For a one-off email, writing your draft and asking AI to polish it works perfectly. When you write the draft, you have already injected persona, context, and tone into the text itself. The AI is simply performing validation.
The four-block prompt structure is intended for another reality: recurring tasks.
That is where the math flips. You write the instruction once, store it as your agent's system prompt (Copilot, ChatGPT, and Claude all support this under different names), and every new interaction inherently executes with that defined behavior.
In Alexandre's words:
“I write this once, store it in my agent's instructions, and every time I interact with it, it automatically uses this format as its baseline.”
Another participant shared a practical workaround: she keeps a pinned conversation specifically for email drafting, with the prompt at the top, appending requests there over time so the model adheres to her pattern.
→ The more recurring the task, the more valuable it is to write the instruction once and never touch it again.
7. Four Details Nobody Mentions
From that same workshop, four practical insights that transform how you operate the technology:
1. Everything turns into tokens, and tokens are what you pay for.
“Innovation with Artificial Intelligence” has 4 words, 40 characters, and 7 tokens. A token is neither a word nor a letter; it is the fundamental compute unit models use to process, bill, and limit text. The context window—the total buffer a conversation can hold—is measured this way, summing your inputs and the model's outputs. Today, it remains the primary boundary across all LLMs.
2. A longer prompt is not a better prompt.
A 2,000-line prompt can yield the exact same result as a 10-line one. And it carries a side effect: in extremely long prompts, AI prioritizes the end and gradually loses focus on the beginning (the 'lost in the middle' phenomenon). If your most critical instruction sits on line 1 of a massive prompt, it may simply be disregarded.
3. Two or three examples are plenty.
Providing one or two reference emails works wonders; providing ten actually hurts, because with too many examples the AI overfits and copies one verbatim rather than generating fresh output adhering to your pattern.
4. Ask for reasoning, not just the final answer.
Asking AI to explain its step-by-step logic forces it to pause and verify whether its output actually answers your original constraints. And it gives you an underutilized leverage point: you get to inspect and correct its reasoning, not just the final text.
Key Takeaway
The quote that summarizes the entire class came from Alexandre himself, and it's what I repeat most when speaking to enterprise leaders:
“The generalist approach will satisfy you up to a certain point. But to solve daily operational pain, you must anchor the model in your specific context. Stay generalist, and you hit a hard ceiling.”
The foundational technology is identical for everyone: the ChatGPT or Copilot you use is the exact same one your competitor uses.
What dictates the outcome is the proprietary context that only you possess. Your workflow, your clients, your writing culture, your department's business rules.
And here lies a truth more valuable than the email itself:
Whoever masters providing context in a conversation is training the exact skill that later matures into agents, automations, and enterprise systems. It is the identical logic—written once, executing thousands of times.
Through the outside lens,
Lucas
PS. In the next edition: what happened when we put Copilot to work live inside Word, Outlook, Excel, and PowerPoint across 24 meetings, 78 emails, and a 5-slide deck. And the two specific places where it failed.
If anyone on your team keeps complaining that “AI gives generic answers”, forward them this article. It is likely the missing mechanism they need.